Accuracy on the Line: On the Strong Correlation Between\n Out-of-Distribution and In-Distribution Generalization

2021/07/09 by J. J. Miller, Miller, John, Rohan Taori +15 · 12 citations
Computer Science · Medicine · #Advanced Neural Network Applications #COVID-19 diagnosis using AI #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Machine Learning and Data Classification

paper · pdf · doi:10.48550/arxiv.2107.04649

openalex publication_date 2021/07/09 · openalex created_date 2021/07/19 · openalex updated_date 2026/07/28

Abstract

For machine learning systems to be reliable, we must understand their\nperformance in unseen, out-of-distribution environments. In this paper, we\nempirically show that out-of-distribution performance is strongly correlated\nwith in-distribution performance for a wide range of models and distribution\nshifts. Specifically, we demonstrate strong correlations between\nin-distribution and out-of-distribution performance on variants of CIFAR-10 &\nImageNet, a synthetic pose estimation task derived from YCB objects, satellite\nimagery classification in FMoW-WILDS, and wildlife classification in\niWildCam-WILDS. The strong correlations hold across model architectures,\nhyperparameters, training set size, and training duration, and are more precise\nthan what is expected from existing domain adaptation theory. To complete the\npicture, we also investigate cases where the correlation is weaker, for\ninstance some synthetic distribution shifts from CIFAR-10-C and the tissue\nclassification dataset Camelyon17-WILDS. Finally, we provide a candidate theory\nbased on a Gaussian data model that shows how changes in the data covariance\narising from distribution shift can affect the observed correlations.\n

Cited by

• Entropy Sentinel: Probing Entropy Traces for LLM Monitoring
• Contrast transfer functions help quantify neural network out-of-distribution generalization in HRTEM
• ConceptScope: Characterizing Dataset Bias via Disentangled Visual Concepts
• Training data composition determines machine learning generalization and biological rule discovery
• Aggregation Hides Out-of-Distribution Generalization Failures from Spurious Correlations
• GaLLoP: Gradient-based Sparse Learning on Low-Magnitude Parameters
• Confidence and Dispersity as Signals: Unsupervised Model Evaluation and Ranking
• Towards Generalizable PDE Dynamics Forecasting via Physics-Guided Invariant Learning
• Scaling Vision-Language Models Is Not Enough to Mitigate Bias
• Sensible Agent: A Framework for Unobtrusive Interaction with Proactive AR Agents
• WILDS: A Benchmark of in-the-Wild Distribution Shifts
• Domain Adaptation via Feature Refinement
• ALSA: Anchors in Logit Space for Out-of-Distribution Accuracy Estimation
• Consensus-Driven Active Model Selection

Related

• Predicting with Confidence on Unseen Distributions
• Robustness May be More Brittle than We Think under Different Degrees of Distribution Shifts
• Empirical Study on Optimizer Selection for Out-of-Distribution Generalization
• The Evolution of Out-of-Distribution Robustness Throughout Fine-Tuning
• Leveraging Unlabeled Data to Predict Out-of-Distribution Performance